+ All Categories
Home > Documents > Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of...

Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of...

Date post: 18-Mar-2020
Category:
Upload: others
View: 3 times
Download: 0 times
Share this document with a friend
13
Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations of the Euler Equations Jared Crean * , Kinshuk Panda * , Anthony Ashley * , and Jason E. Hicken Rensselaer Polytechnic Institute, Troy, New York, 12180 We present an extensible Julia-based solver for the Euler equations that uses a summation- by-parts (SBP) discretization on unstructured triangular grids. While SBP operators have been used for tensor-product discretizations for some time, they have only recently been extended to simplices. Here we investigate the accuracy and stability properties of simplex- based SBP discretizations of the Euler equations. Non-linear stabilization is a particular concern in this context, because SBP operators are nearly skew-symmetric. We consider an edge-based stabilization method, which has previously been used for advection-diffusion- reaction problems and the Oseen equations, and apply it to the Euler equations. Addi- tionally, we discuss how the development of our software has been facilitated by the use of Julia, a new, fast, dynamic programming language designed for technical computing. By taking advantage of Julia’s unique capabilities, code that is both efficient and generic can be written, enhancing the extensibility of the solver. I. Introduction High-order accurate discretizations have the potential to be more efficient than low-order methods in the numerical solution of partial differential equations [1, 2]. Among such methods, summation-by-parts (SBP) operators [3] are an attractive option for computational fluid dynamics (CFD), because they are high- order accurate and time stable for linear equations. However, classical SBP operators are one-dimensional finite-difference methods, so most of the previous work using high-order SBP operators has been limited to tensor-product discretizations on structured or multi-block grids [39]. Recently, Hicken, Del Rey Fern´ andez, and Zingg [10] proposed a theoretical framework for multi- dimensional SBP operators that is suitable for discretizations on unstructured grids. The framework was illustrated by constructing SBP operators for triangular and tetrahedral elements. These simplex SBP operators produce discretizations that are similar to the mass-lumped spectral-element method [11, 12]. SBP-based discretizations, while time stable, may still suffer from aliasing errors that arise in the dis- cretization of non-linear partial differential equations (PDEs). Stabilization methods to control these errors have been studied extensively over the last half century, particularly in the context of advection-dominated problems [1317]. In this work, we focus on continuous SBP discretizations and draw upon stabilizations suitable for continuous Galerkin (CG) methods that are high-order accurate and locally conservative [18, 19]. Among the most popular stabilizations for finite-element methods is the streamwise-upwind Petrov- Galerkin (SUPG) method, which established the effectiveness and accuracy of artificial dissipation when consistency is maintained [14]. Although popular, it suffers from several limitations including non-physical coupling of velocity and pressure in the stabilization term and dual inconsistency [20]. * Graduate Student, Department of Mechanical, Aerospace, and Nuclear Engineering, Student Member AIAA Assistant Professor, Department of Mechanical, Aerospace, and Nuclear Engineering, Member AIAA 1 Downloaded by Jason Hicken on February 8, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2016-1328 54th AIAA Aerospace Sciences Meeting 4-8 January 2016, San Diego, California, USA AIAA 2016-1328 Copyright © 2015 by Jared Crean, Kinshuk Panda, Anthony Ashley, Jason E. Hicken. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission. AIAA SciTech
Transcript
Page 1: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

Investigation of Stabilization Methods for

Multi-Dimensional Summation-by-parts

Discretizations of the Euler Equations

Jared Crean∗, Kinshuk Panda∗, Anthony Ashley∗, and Jason E. Hicken†

Rensselaer Polytechnic Institute, Troy, New York, 12180

We present an extensible Julia-based solver for the Euler equations that uses a summation-by-parts (SBP) discretization on unstructured triangular grids. While SBP operators havebeen used for tensor-product discretizations for some time, they have only recently beenextended to simplices. Here we investigate the accuracy and stability properties of simplex-based SBP discretizations of the Euler equations. Non-linear stabilization is a particularconcern in this context, because SBP operators are nearly skew-symmetric. We consider anedge-based stabilization method, which has previously been used for advection-diffusion-reaction problems and the Oseen equations, and apply it to the Euler equations. Addi-tionally, we discuss how the development of our software has been facilitated by the use ofJulia, a new, fast, dynamic programming language designed for technical computing. Bytaking advantage of Julia’s unique capabilities, code that is both efficient and generic canbe written, enhancing the extensibility of the solver.

I. Introduction

High-order accurate discretizations have the potential to be more efficient than low-order methods inthe numerical solution of partial differential equations [1, 2]. Among such methods, summation-by-parts(SBP) operators [3] are an attractive option for computational fluid dynamics (CFD), because they are high-order accurate and time stable for linear equations. However, classical SBP operators are one-dimensionalfinite-difference methods, so most of the previous work using high-order SBP operators has been limited totensor-product discretizations on structured or multi-block grids [3–9].

Recently, Hicken, Del Rey Fernandez, and Zingg [10] proposed a theoretical framework for multi-dimensional SBP operators that is suitable for discretizations on unstructured grids. The framework wasillustrated by constructing SBP operators for triangular and tetrahedral elements. These simplex SBPoperators produce discretizations that are similar to the mass-lumped spectral-element method [11,12].

SBP-based discretizations, while time stable, may still suffer from aliasing errors that arise in the dis-cretization of non-linear partial differential equations (PDEs). Stabilization methods to control these errorshave been studied extensively over the last half century, particularly in the context of advection-dominatedproblems [13–17]. In this work, we focus on continuous SBP discretizations and draw upon stabilizationssuitable for continuous Galerkin (CG) methods that are high-order accurate and locally conservative [18,19].

Among the most popular stabilizations for finite-element methods is the streamwise-upwind Petrov-Galerkin (SUPG) method, which established the effectiveness and accuracy of artificial dissipation whenconsistency is maintained [14]. Although popular, it suffers from several limitations including non-physicalcoupling of velocity and pressure in the stabilization term and dual inconsistency [20].

∗Graduate Student, Department of Mechanical, Aerospace, and Nuclear Engineering, Student Member AIAA†Assistant Professor, Department of Mechanical, Aerospace, and Nuclear Engineering, Member AIAA

1

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

54th AIAA Aerospace Sciences Meeting

4-8 January 2016, San Diego, California, USA

AIAA 2016-1328

Copyright © 2015 by Jared Crean, Kinshuk Panda, Anthony Ashley, Jason E. Hicken. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.

AIAA SciTech

Page 2: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

Instead, we investigate a stabilization based on penalizing jumps in the gradient, which has been shownto be effective, and in some cases optimal, in stabilizing advection-dominated problems [16]. The gradient-jump penalty of Burman and Hansbo was shown to bound the energy norm of the discrete solution of theconvection-diffusion-reaction and Oseen equations [21,22], and we adapt it here to the Euler equations.

In addition to studying stabilization of multi-dimensional SBP methods, one of our motivations for thepresent work is to develop a versatile multi-physics code for PDE-constrained optimization. PDE solvers usedfor optimization have several requirements beyond traditional analysis codes, most importantly the abilityto calculate derivatives. To this end, we have implemented our solver in a new, fast, dynamic programminglanguage called Julia, which enables the use of abstraction while retaining computational efficiency. Weelaborate on the Julia implementation further below.

The remainder of the paper is organized as follows. The SBP discretization is discussed in Section II,and Section III describes edge stabilization. The use of abstractions in Julia and the computational benefitsare detailed in Section IV. Numerical results for both steady and unsteady cases are given in Section V, andconclusions are provided in Section VI.

II. Summation-by-parts Discretization of the Euler Equations

Our discretizations use the multi-dimensional SBP simplex operators recently proposed in Ref. [10]. Tokeep the presentation self contained, the definition of a two-dimensional SBP first-derivative operator in theξ direction is provided below for a generic reference element Ωe. The definition for the derivative operator inthe η direction is analogous and uses the same norm/mass matrix H. In the definition, we make use of themonomials Pk(ξ, η) ≡ ξiηj−i, where i, j, and k are related by k = j(j + 1)/2 + i+ 1, for all j ∈ 0, 1, . . . , pand i ∈ 0, 1, . . . , j.

Definition 1. Two-dimensional summation-by-parts operator: Consider an open and bounded do-main Ωe ⊂ R2 with a piecewise-smooth boundary Γe. The matrix Dξ ∈ Rn×n is a degree p SBP approximationto the first derivative ∂

∂ξ on the nodes S = (ξi, ηi)ni=1 if

1. Dξpk = p′k, ∀ k ∈ 1, 2, . . . , (p+ 1)(p+ 2)/2,where pk ∈ Rn and p′k ∈ Rn denote Pk and ∂Pk/∂ξ, respectively, evaluated at the nodes in S;

2. Dξ = H−1Sξ, where H is symmetric positive-definite, and;

3. Sξ = Qξ + 12 Eξ, where QT

ξ = −Qξ, ETξ = Eξ, and Eξ satisfies

pTk Eξpm =

∮ΓePkPmnξdΓ, ∀ k,m ∈ 1, 2, . . . , (τ + 1)(τ + 2)/2,

where τ ≥ p, and nξ is the ξ component of n = [nξ, nη]T

, the outward pointing unit normal on Γe.

The simplex-based SBP operators that we consider in this work were constructed with diagonal norm/massmatrices whose entries define a cubature rule with positive weights. Unlike most finite-difference methods,SBP discretizations approximate the weak form when integrated using their mass matrix. Unlike finite-element methods, SBP methods do not have unique shape functions. They only specify basis values andderivatives at the nodes (property 1 above) and require the matrix operators to obey properties 2 and 3.

In the following sections, we illustrate how multi-dimensional SBP operators are used to discretization theEuler equations, and we highlight the close connection between SBP discretizations and the finite-elementmethod.

A. Conservative Variable Formulation

The two-dimensional Euler equations are discretized in space using SBP operators on a simplex mesh. Theequations in conservation form are

∂q

∂t+∇ · F = S, (1)

2

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 3: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

where q = [ρ, ρu, ρv, E]T

denotes the conservative variables, S is the source term, and the Euler fluxes are

F =

ρu ρv

ρu2 + p ρuv

ρuv ρv2 + p

(E + p)u (E + p)v

=[Fx Fy

].

The calorically perfect ideal gas law is used to close the system. The semi-linear weak form of the Eulerequations is derived below, in order to illustrate some particular details. Readers familiar with the derivationcan proceed directly to Equation (3).

Consider an arbitrary element Ω with boundary Γ, and let x = (x(ξ, η), y(ξ, η)) be a mapping from

the reference element Ωe to Ω, where (ξ, η) are the reference element coordinatesa. Let J =(∂ξ∂x

)be the

mapping Jacobian, and let |J | denote its determinant. Transforming the fluxes in Equation (1) to referencespace and multiplying the entire equation by 1

|J| , we obtain

1

|J |∂q

∂t+∇ξ · Fξ =

S

|J |, (2)

where the transformed fluxes are

Fξ ≡1

|J |

(Fx

∂ξ

∂x+ Fy

∂ξ

∂y

), and Fη ≡

1

|J |

(Fx

∂η

∂x+ Fy

∂η

∂y

).

If the metric invariants are satisfied — they are for the affine transformations considered here — the 1|J|

factor can be moved inside the divergence operator to maintain conservative form in reference space [23].

Next, we introduce a weighting function w ∈ [V]4, where the weighting space V is an appropriate Hilbert

space. Multiplying element-wise by the vector w, integrating over the reference element Ωe, and applyingintegration by parts to the flux term, yields∫

Ωew∂q

∂t

1

|J |dΩe =

∫ΩewS

1

|J |dΩe +

∫Ωe∇ξw · [Fξ,Fη] dΩe −

∫Γew [Fξ,Fη] dΓ,

where dΓ = ndΓ and n = (nξ, nη) is the outward unit vector normal to the boundary Γe.On each element we introduce a finite dimensional approximation q ≈ Nb(ξ)qb ∈ δh, where δh is the trial

space. In addition, the non-linear fluxes are projected onto the finite-dimensional space via∫Ωew[Nb(ξ)

(Fξ

)b− Fξ(Nb(ξ)qb)

]dΩe = 0.

A similar projection is used to define Sb. Note that all quantities with hats are coefficients that are functionsof time only, and all spatial dependence is contained in the shape functions Nb(ξ). The subscript b indicatesthe basis function index, and repeated indices are summed.

Using the Bubnov-Galerkin approach, we can express w = Nawa ∈[Vh]4

, and therefore δh = Vh.Requiring the weak-form to be satisfied for all choices of wa, results in∫

ΩeNaNb

∂qb∂t

1

|J |dΩe =

∫ΩeNaNbSb

1

|J |dΩe +

∫Ωe∇ξNa ·

[Nb

(Fξ

)b, Nb

(Fη

)b

]dΩe

−∫

ΓeNaNb

[(Gξ

)b,(Gη

)b

]dΓ ∀ Na ∈ Vh.

(3)

In order to impose the boundary conditions weakly, the fluxes in the boundary integrals have been replacedwith numerical flux functions, specifically Roe flux functions, denoted Gξ; see, for example, [24].

aIn this work, we consider only affine mappings between reference elements and physical elements.

3

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 4: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

We now replace all operations on shape functions with the SBP matrix operators. Defining J−1 ≡diag(1/|J1|, 1/|J2|, . . . , 1/|Jn|), we can replace the first integral on the right hand side of Equation (3) with

HJ−1S, where H is the diagonal mass matrix. The second term is the volume integral contribution to theelement stiffness matrix, which is approximated by STξ = ∂Na

∂ξ Nb [10]. Note that the element stiffness matrices

in the parametric coordinate directions can be expressed as STξ = (HDξ)T

; see Definition 1. As a result,a single matrix multiplication performs the action of integrating the stiffness contribution over an element.The final integral in Equation (3) can be expressed using the boundary integration matrices, Eξ and Eη.

Applying these simplifications to the weak form gives

HJ−1 ∂q

∂t= HJ−1S + STξ Fξ + STη Fη − EξGξ − EηGη. (4)

The SBP weak form (4) can be rearranged into its conventional finite-difference strong form by multiplyingfrom the left by the inverse norm matrix, H−1:

∂t

(J−1q

)= J−1S + H−1

[STξ Fξ + STη Fη − EξGξ − EηGη

]= J−1S − DξFξ − DηFη − H−1Eξ

[Gξ − Fξ

]− H−1Eη

[Gη − Fη

],

where we have used properties 2 and 3 from Definition 1. Note that the two terms multiplied by H−1

on the right-hand side, the so-called simultaneous approximation terms, represent penalties with vanishingtruncation errors that impose the boundary conditions weakly [25,26].

B. Entropy Variable Formulation

The formulation above uses SBP operators to discretize the Euler equations based on the conservativevariables. The resulting semi-discrete system is not entropy stable and may not be suitable for long-timesimulations. On the other hand, Hughes et al. proved that a Galerkin discretization of the symmetrized Eulerand Navier-Stokes equations in entropy variables would satisfy the discrete Clausis-Duhem inequality [27,28]. Entropy stability has been shown to improve the robustness of high-order discretizations involvingturbulence [29].

Motivated by the results in [29], we consider an SBP discretization of the symmetrized Euler equations inorder to take advantage of the proven “energy” stability in the physically significant entropy norm. However,in this preliminary work we have not implemented a skew-symmetric discretization of the derivative, so oursemi-discrete scheme is not provably entropy stable. Provable entropy stability will be the focus of futurework.

We discretize the entropy variable formulation using SBP operators starting with Equation (2). Intro-ducing a change of variables from q to v, where v are the so-called entropy variables, defined in [28], wehave

∂q

∂t=∂q

∂v

∂v

∂t= A0

∂v

∂t,

and

∇ · F =∂Fi∂xi

=∂Fi∂q

∂q

∂xi= Ai

∂q

∂xi= ASi

∂q

∂v

∂v

∂xi= ASi A0

∂v

∂xi= ASi

∂v

∂xi= ∇ · F S

where F S are the Euler fluxes and the ASi are the flux Jacobians with respect to entropy variables. Thedefinitions of these quantities in terms of entropy variables are given in the appendix of [28]. Thus the Eulerequations become

A0∂v

∂t+∇ · F S = S. (5)

The derivation of the SBP weak form follows in the same fashion presented in Subsection II.A. The result is

HJ−1

(A∂v

∂t

)= HJ−1S + STξ F

Sξ + STη F

Sη − EξG

Sξ − EηG

Sη , (6)

4

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 5: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

where A is a block-diagonal matrix holding the A0 matrices evaluated at the nodes.

III. Edge Stabilization

As discussed in the introduction, SBP operators do not provide the dissipation necessary to preventaliasing errors. To address this, we consider the edge stabilization originally presented by Douglas andDupont [30] (see also [21] and [16]). It is obtained by adding the term

J(w, q) = −1

2

∑ν

∫Γν

γstabh2ν [∇q] · [∇w] dΓ = −1

2

∑ν

∫Γν

γstabh2ν [n · ∇q] [n · ∇w] dΓ

to Equation (3), or the entropy-variable equivalent, where Γν denotes an interior edge, hν is the nominal edgelength, and γstab is a chosen constant. Square brackets denote the jump operator on the element boundary:[u] = limδ→0 u(x + δn) − u(x − δn) for x ∈ Γν . The jump-stabilization term acts at the interface betweentwo elements, penalizing differences in the normal derivative.

Our solver is intended for compressible flow problems, so we have adapted the original edge-stabilizationmethod. In particular, we use the spectral radius of the normal flux Jacobian as a scaling term to ensure thestabilization is dimensionally consistent with the Euler equations, and we consider the jump in all solutionvariables across the element boundaries. The adapted form of the stabilization term used in the solver canbe seen in Equation (7).

J(w, q) = −1

2

∑ν

∫Γν

(|u · n| + a)γstabh2ν [n · ∇q] [n · ∇w] dΓ, (7)

where a is the local speed of sound. We will explore the effects of γstab in Section V.To implement edge stabilization within an SBP discretization, we use the following approximation to the

normal-derivative jump operator.

[n · ∇] ≈ Dν ≡ (Bξ,L,νLνDξ,L + Bη,L,νLνDη,L) + (Bξ,R,νRνDξ,R + Bη,R,νRνDη,R) .

Here, Dξ,L and Dη,L denote the SBP derivative operators on the nominal left side of the interface, and Dξ,R

and Dη,R denote the corresponding operators on the right side of the interface. The operators Lν and Rν arebinary matrices that select the boundary nodes on Γν for the left and right elements, respectively; Lν and Rνare nν ×n rectangular matrices, where nν is the number of nodes on Γν . The matrices Bξ,L,ν , Bη,L,ν , Bξ,R,ν ,and Bη,L,ν contain geometric information that accounts for the transformation from reference to physicalspace. They are nν × nν diagonal matrices given by

Bξ,L,ν = diag [(∇ξ · ∇ξ)nξ + (∇ξ · ∇η)nη]L , Bη,L,ν = diag [(∇ξ · ∇η)nξ + (∇η · ∇η)nη]L ,

Bξ,R,ν = diag [(∇ξ · ∇ξ)nξ + (∇ξ · ∇η)nη]R , Bη,R,ν = diag [(∇ξ · ∇η)nξ + (∇η · ∇η)nη]R ,

where (nξ, nη) is the normal to Γν in reference space.Using the edge mass matrix Mν to approximate integration over Γν , the SBP form of edge stabilization

isJh(wi, qj) = −γ

2

∑ν

h2ν (Dνwi)

T MνΥDνqj ,

where wi, qj ∈ Rn for i, j = 1 : 4 are components of thew and q vectors at the nodes, and Υ = diag(|u·n|+a)contains the fastest acoustic wave speeds at each node on Γν . For SBP discretizations, the test function wiis a binary unit vector taking the value of 1 for one node and zero otherwise. Thus, by considering eachnode in turn, the above form becomes the vector

Jh(I, qj) = −γ2

∑ν

h2νDT

ν MνΥDνqj ,

5

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 6: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

where I is the n× n identity.If |u ·n|+ a is constant over Γν , then edge stabilization does not interfere with entropy stability, because

the operator is negative semi-definite. Proving entropy stability in the case of a spatially varying |u · n|+ ais also possible, but requires interpolation to cubature nodes on Γν ; see, for example, [31].

IV. Solver Abstraction and Implementation

A. Implementation in Julia

Having described the SBP discretization, we now pause to discuss our implementation and solution of theresulting non-linear algebraic equations using the programming language Julia. One of the main features ofJulia is a generic, type optional, programming style where type information can be specified by the user orinferred by the compiler. When a function is called with a particular set of values as arguments, the JustIn Time (JIT) compiler uses the types of the values to compile a version of the function specialized to thosetypes, as if the programmer had specified the datatype of every variable. This gives Julia the efficiencyof statically typed languages like C (without vectorization), but with the flexibility to call functions withdifferent argument types [32], and permits the use of abstraction in a natural and readable syntax, even inlow-level numerical routines.

The evaluation of the Euler equations in the form of Equation (4) is implemented as a system of non-linear algebraic equations. Using Julia’s parametric type system, the entire calculation is parameterized onthe types of the input variables q, the type of the result ∂q

∂t , and the type of the mesh Jacobian terms ∂ξ∂x .

As a result, it is possible to differentiate with respect to either the solution variables or the mesh variables.One use of the parameterization is to calculate the Jacobian for Newton’s method, which we use to solve

steady problems. An abstract form of algorithmic differentiation (AD) is implemented, taking in a seedvector whose elements can be of any datatype. Because of Julia’s type inference capability and the JITcompiler, when the evaluation function is called with complex-perturbed solution variables, every function iscompiled with datatypes of all the variables known to the compiler. If a different AD datatype is used, suchas hyper-dual numbers [33], all the numerical routines are recompiled, generating efficient machine code tooperate on dual numbers. The result is a highly efficient mechanism for algorithmic differentiation, with thefull range of compiler optimization available and no source code duplication.

One of the benefits of the parameterization is its extensibility. For example, in order to implement entropyvariables, an additional type parameter is introduced. This parameter enables the use of Julia’s multipledispatch system to call the correct methods for the variables being used. For example, methods to calculatepressure and the Euler flux at a node are defined for both variables, and the compiler chooses between thembased on the type parameter during method dispatch at compile time. Thus, the implementation of entropyvariables required only defining the node-based operations and the transformation matrices described inSection II.B. This ease of extensibility for low-level routines demonstrates the utility of the multiple dispatchparadigm in designing numerical software.

B. Algebraic Equation Solvers

Steady flow problems are solved using Newton’s method. To solve the linear systems that arise in Newton’smethod, a sparse multi-frontal LU factorization from UMFPACK is employed [34]. For unsteady problems,we use the classical 4th-order explicit Runge-Kutta method.

In order to calculate the Jacobian for Newton’s method with as few (complex) residual evaluations aspossible, mesh coloring and an element-based data structure are used. In the element-based data arrays, thesolution and residual values are stored for every node on each element, duplicating the values for nodes thatare shared between elements. This enables perturbation of the q variables on one element without affectingneighboring elements.

For the unstabilized equations a distance-0 coloring (i.e. all elements are the same color) is sufficientto ensure there is a one-to-one mapping from residual perturbations to perturbations in q. Each degree offreedom on an element must be perturbed independently, therefore the Jacobian can be calculated in m = 4n

6

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 7: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

(a) Mesh (b) Contours of ρ

Figure 1: Steady isentropic vortex solution

residual evaluations, where m is the number of degrees of freedom on an element.For edge stabilization, where perturbations to a degree of freedom affect all neighboring elements that

share an edge, a distance-2 graph coloring is necessary, where graph vertices correspond to mesh elementsand graph edges connect elements that share a mesh edge. We use Algorithm 3.1 from Ref. [35] to performthis coloring, which has an upper bound of ∆2 + 1 colors, where ∆ is the maximum number of neighboringelements. Considering only edge neighbors for triangular elements, the maximum number of colors is 10,and the maximum number of residual evaluations is 10m. The graph-based Parallel Unstructured MeshInterface [36] was used for all meshing, and it facilitated the implementation of the coloring algorithm byallowing querying of the topological adjacencies of mesh elements.

V. Results

A. Isentropic Vortex

An isentropic vortex problem was used to verify the asymptotic convergence rates of the discretizations. Thevortex problem consists of a quarter circle domain with an inner radius rin = 1 and outer radius rout = 3.An example of the expected computational solution can be seen in Figure 1(b), in which density is plotted.Here, the problem was solved with degree p = 2 elements, with 20 elements along each boundary edge; seeFigure 1(a) for the corresponding mesh.

The analytical solution of this problem is known to be

ρ(r) = ρin

[1 +

γ − 1

2M2

in

(1− r2

in

r2

)] 1γ−1

.

We have set the inner-radius density to ρin = 2 and inner-radius Mach number to Min = 0.95. The specificheat ratio γ = 1.4 is used for all problems. All other flow variables can be calculated using the isentropicrelations. Note that the analytical solution is imposed (weakly) along all boundaries.

To determine the convergence rates of the discretizations, the steady vortex problem was solved on asequence of meshes, using the analytic solution as the initial guess. Newton’s method was run until anabsolute residual tolerance of 10−12 was achieved, or the residual was reduced by ten orders relative to theresidual for a uniform flow. The SBP approximation to the continuous L2 norm,

√uTHJu, is used for all

error-norm calculations.Figure 2 shows the convergence rates for the unstabilized and edge-stabilized discretizations. As in [10]

for the linear advection equation, the p = 2 and p = 4 discretizations show suboptimal convergence rates of

7

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 8: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

2e-2 6e-2 1e-1h

10-7

10-6

10-5

10-4

10-3

10-2

10-1

L2 e

rror

1.8:1

2:1

3.8:1

p = 1p = 2p = 3p = 4

(a) No stabilization

2e-2 6e-2 1e-1h

10-7

10-6

10-5

10-4

10-3

10-2

10-1

L2 e

rror

1.9:1

3:1

3.7:1

4.4:1p = 1p = 2p = 3p = 4

(b) Edge stabilization

Figure 2: Effect of mesh refinement upon L2 error

Degree Convergence: No Stabilization Convergence: Edge Stabilization γstab

p = 1 1.8 1.9 0.01

p = 2 2.0 3.0 0.9

p = 3 - 3.7 7.35

p = 4 3.8 4.4 25.0

Table 1: Convergence rates observed in the steady-vortex problem

8

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 9: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

order p. The odd-order discretizations show oscillatory convergence, especially the p = 3 discretization. Incontrast, the edge-stabilized discretizations exhibit well-behaved convergence.

For edge-stabilization, values of the stabilization parameter that produce optimal p+1 convergence rateswere easily found for the p = 1 and p = 2 elements. A range of values was found to produce optimalconvergence rates, so the value corresponding to the minimal error norm is reported here. For the higherorder elements, a Golden-Section optimization was performed to find the parameter value that maximizedconvergence rate. Despite this optimization, the p = 3 and p = 4 elements exhibit sub-optimal convergencerates, although the p = 2 and p = 4 convergence rates are improved relative to the unstabilized case. Theconvergence rates are estimated using a least-squares analysis of the five finest meshes and are listed inTable 1.

B. Unsteady Vortex

We now investigate the conservative- and entropy-variable formulations on an unsteady problem with andwithout edge stabilization. Specifically, we consider the unsteady isentropic vortex problem; see, for exam-ple, [37]. The analytical solution is known to be [38]

u =1− εy

2πexp

(f(x, y, t)

2

), v =

ε((x− x0)− t)2π

exp

(f(x, y, t)

2

)ρ =

(1− ε2(γ − 1)M2

8π2exp (f(x, y, t))

) 11−γ

, p =ργ

γM2, (8)

where f(x, y, t) = 1− (((x− x0)− t)2 + y2), the Mach number is set to M = 0.5, ε, the vortex strength, isset to 1, and x0, the x coordinate of the center of the vortex at t = 0, is 5. The y coordinate of the vortex’scenter is zero. We solve the problem on a rectangular domain x ∈ [0, 20] and y ∈ [−5, 5]. The analyticalsolution is imposed for both the initial condition and boundary conditions. The simulation was run to amaximum time of 2 seconds.

A mesh defined by 50 elements along each edge was used for the p = 1 discretization, correspondingto 10,404 degrees-of-freedom. For the higher order elements, meshes with the closest number of degrees-of-freedom possible were used, subject to the constraint that the number of elements along each edge is equal;see Table 2.

The maximum stable CFL number was determined to be 0.0001 for the p = 4 stabilized discretizationand was used for all runs. The use of edge stabilization significantly reduced the maximum stable CFLnumber compared to the unstabilized case. In calculating the CFL number, the minimum distance betweennodes for each degree element was used as the mesh spacing.

For the p = 1 discretization, the same stabilization constant was used as for the steady vortex problem.For high-order elements, γstab = 0.5 proved to be sufficient to maintain stability. For reasons of computationaleconomy, higher values of γstab were not studied. Table 2 lists the value of the edge stabilization parameterγstab used for all degree operators.

Degree CFL γstab DOFs

p = 1 0.0001 0.01 10,404

p = 2 0.0001 0.5 10,560

p = 3 0.0001 0.5 10,804

p = 4 0.0001 0.5 10,644

Table 2: CFL and edge stabilization parameter values used for the unsteady vortex problem. The p = 4stabilized elements dictated the maximum stable timestep used for all discretizations.

Figure 3 shows the solution field at the final time. Density is plotted for the p = 2 discretization with andwithout stabilization. Qualitatively, one can see the effects of adding stabilization; Figure 3(a) demonstrates

9

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 10: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

(a) p = 2, no stabilization (b) p = 2, stabilized

Figure 3: Density at final time of the unsteady vortex problem

slight oscillations around the vortex that indicate instability, whereas Figure 3(b) shows significantly feweroscillations.

Because the exact solution is isentropic, it is desirable for a numerical method to retain this property.The results in Figure 4 indicate that both the conservative and entropy formulations increase entropy overtime. The unstabilized formulations generate similar amounts of entropy, but for p = 1 through p = 3the stabilized entropy variable formulation generates less entropy than the stabilized conservative formula-tion. This suggests the entropy formulation is effective in reducing non-physical entropy generated by thestabilization.

While increasing entropy is expected for the unstabilized conservative case, Hughes et al. proved that thefinite-element discretization of the Euler equations in entropy variables will inherit the continuous equation’sproperty of conserving entropy [28]. As noted in Section II.B, a skew-symmetric discretization is necessaryto realize provable entropy stability and will be the focus of future work.

VI. Conclusion

High-order simplex SBP methods show promise as a means to efficiently solve PDEs numerically onunstructured grids; however, they require stabilization for non-linear problems. To this end, we have in-vestigated edge stabilization for simplex SBP operators, examining both the convergence rate and entropystability of the spatial discretizations. Retaining high-order convergence in the stabilized scheme is neces-sary in order to realize the computational efficiency of high-order methods. Additionally, entropy stabilityis important for long-time simulations, such as those used for unsteady turbulence simulations.

We have shown that edge stabilization is effective in stabilizing a simplex-based SBP discretization of theEuler equations for all orders of simplex SBP elements, although it significantly increases the cost of usingalgorithmic differentiation to compute the Jacobian and, more significantly, imposes a restriction on theCFL number for unsteady problems using explicit time marching. For unsteady problems, entropy variablesincrease the solution accuracy of the stabilized scheme with respect to entropy, but does not render thediscretization entropy stable without a skew-symmetric derivative approximation. Further investigation isrequired into a skew-symmetric SBP formulation for entropy stability.

Acknowledgements

A. Ashley and J. Hicken gratefully acknowledge the support of the Air Force Office of Scientific ResearchAward FA9550-15-1-0242 under Dr. Jean-Luc Cambier.

10

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 11: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

0.0 0.5 1.0 1.5 2.0t

-10-2

-10-3

-10-4

-10-5

-10-6

0

10-6

10-5

10-4

10-3

∆S

Cons. Vars.Entropy Vars.Cons. Vars., stab.Entropy Vars., stab.

(a) p = 1

0.0 0.5 1.0 1.5 2.0t

-10-2

-10-3

-10-4

-10-5

-10-6

0

10-6

10-5

10-4

10-3

∆S

Cons. Vars.Entropy Vars.Cons. Vars., stab.Entropy Vars., stab.

(b) p = 2

0.0 0.5 1.0 1.5 2.0t

-10-2

-10-3

-10-4

-10-5

-10-6

0

10-6

10-5

10-4

10-3

∆S

Cons. Vars.Entropy Vars.Cons. Vars., stab.Entropy Vars., stab.

(c) p = 3

0.0 0.5 1.0 1.5 2.0t

-10-2

-10-3

-10-4

-10-5

-10-6

0

10-6

10-5

10-4

10-3

∆S

Cons. Vars.Entropy Vars.Cons. Vars., stab.Entropy Vars., stab.

(d) p = 4

Figure 4: Change in entropy over solution time

11

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 12: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

References

1Kreiss, H.-O. and Oliger, J., “Comparison of accurate methods for the integration of hyperbolic equations,” Tellus,Vol. 24, No. 3, 1972, pp. 199–215.

2Swartz, B. and Wendroff, B., “The Relative Efficiency of Finite Difference and Finite Element Methods. I: HyperbolicProblems and Splines,” SIAM Journal on Numerical Analysis, Vol. 11, No. 5, 1974, pp. pp. 979–993.

3Kreiss, H.-O. and Scherer, G., “Finite element and finite difference methods for hyperbolic partial differential equations,”Mathematical Aspects of Finite Elements in Partial Differential Equations, edited by C. de Boor, Mathematics Research Center,the University of Wisconsin, Academic Press, 1974.

4Strand, B., “Summation by Parts for Finite Difference Approximations for d/dx,” Journal of Computational Physics,Vol. 110, No. 1, 1994, pp. 47 – 67.

5Hicken, J. E. and Zingg, D. W., “A parallel Newton-Krylov solver for the Euler equations discretized using simultaneousapproximation terms,” AIAA Journal , Vol. 46, No. 11, Nov. 2008, pp. 2773–2786.

6Nordstrom, J., Gong, J., van der Weide, E., and Svard, M., “A stable and conservative high order multi-block methodfor the compressible Navier–Stokes equations,” Journal of Computational Physics, Vol. 228, No. 24, 2009, pp. 9020–9035.

7Mattsson, K. and Carpenter, M. H., “Stable and accurate interpolation operators for high-order multiblock finite differencemethods,” SIAM Journal on Scientific Computing, Vol. 32, No. 4, 2010, pp. 2298–2320.

8Svard, M. and Nordstrom, J., “Review of summation-by-parts schemes for initial-boundary-value-problems,” Journal ofComputational Physics, Vol. 268, No. 1, 2014, pp. 17–38.

9Del Rey Fernandez, D. C., Hicken, J. E., and Zingg, D. W., “Review of summation-by-parts operators with simultaneousapproximation terms for the numerical solution of partial differential equations,” Computers and Fluids, 2014.

10Hicken, J. E., Del Rey Fernandez, D. C., and Zingg, D. W., “Multidimensional summation-by-parts operators: generaltheory and application to simplex elements,” SIAM Journal on Scientific Computing, September 2015, Submitted (in revision).

11Cohen, G., Joly, P., Roberts, J. E., and Tordjman, N., “Higher Order Triangular Finite Elements with Mass Lumpingfor the Wave Equation,” SIAM Journal on Numerical Analysis, Vol. 38, No. 6, 2001, pp. pp. 2047–2078.

12Giraldo, F. X. and Taylor, M. A., “A diagonal-mass-matrix triangular-spectral-element method based on cubature points,”Journal of Engineering Mathematics, Vol. 56, No. 3, 2006, pp. 307–322.

13Von Neumann, J. and Richtmyer, R. D., “A Method for the Numerical Calculation of Hydrodynamic Shocks,” Journalof Applied Physics, Vol. 21, No. 3, 1950.

14Brooks, A. N. and Hughes, T. J. R., “Streamline upwind/Petrov-Galerkin formulations for convection dominated flowswith particular emphasis on the incompressible Navier-Stokes equations,” Computer Methods in Applied Mechanics and Engi-neering, Vol. 32, No. 1–3, 1982, pp. 199 – 259.

15Rice, J. G. and Schnipke, R. J., “A monotone streamline upwind finite element method for convection-dominated flows,”Computer Methods in Applied Mechanics and Engineering, Vol. 48, No. 3, 1985, pp. 313 – 327.

16Burman, E., Fernandez, M. A., and Hansbo, P., “Continuous Interior Penalty Finite Element Method for Oseen’s Equa-tions,” SIAM Journal on Numerical Analysis, Vol. 44, No. 3, 2006, pp. pp. 1248–1274.

17Hughes, T. J. R. and Brooks, A., “A multidimensional upwind scheme with no crosswind diffusion,” Finite elementmethods for convection dominated flows, Vol. 34, 1979, pp. 19–35.

18Hughes, T. J. R., Engel, G., Mazzei, L., and Larson, M. G., “The Continuous Galerkin Method Is Locally Conservative,”Journal of Computational Physics, Vol. 163, No. 2, 2000, pp. 467–488.

19Venkatakrishnan, V., Allmaras, S., Kamenetskii, D., and Johnson, F., chap. Higher Order Schemes for the CompressibleNavier-Stokes Equations, Fluid Dynamics and Co-located Conferences, American Institute of Aeronautics and Astronautics,Jun 2003.

20Collis, S. S. and Heinkenschloss, M., “Analysis of the streamline upwind/Petrov Galerkin method applied to the solutionof optimal control problems,” Tech. Rep. TR02-01, Houston, Texas, 2002.

21Burman, E. and Hansbo, P., “Edge stabilization for Galerkin approximations of convection–diffusion–reaction problems,”Computer Methods in Applied Mechanics and Engineering, Vol. 193, No. 15–16, 2004, pp. 1437–1453, Recent Advances inStabilized and Multiscale Finite Element Methods.

22Burman, E., Fernandez, M. A., and Hansbo, P., “Edge stabilization for the incompressible Navier-Stokes equations: acontinuous interior penalty method,” 2004, CMCS-ARTICLE-2004-006.

23Pulliam, T. H., “Efficient Solution Methods for the Navier-Stokes Equations,” Lecture Notes for the Von Karman Institutefor Fluid Dynamics Lecture Series.

24Laney, C., Computational Gas Dynamics, Cambridge University Press, 1998, pp. 71–105.25Carpenter, M. H., Gottlieb, D., and Abarbanel, S., “Time-stable boundary conditions for finite-difference schemes solving

hyperbolic systems: Methodology and application to high-order compact schemes,” Journal of Computational Physics, Vol. 111,No. 2, 1994, pp. 220–236.

26Carpenter, M. H., Nordstrom, J., and Gottlieb, D., “Revisiting and Extending Interface Penalties for Multi-domainSummation-by-Parts Operators,” Journal of Scientific Computing, Vol. 45, No. 1-3, 2010, pp. 118–150.

27Harten, A., “On the symmetric form of systems of conservation laws with entropy,” Journal of Computational Physics,Vol. 49, No. 1, 1983, pp. 151–164.

12

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28

Page 13: Investigation of Stabilization Methods for Multi-Dimensional … · 2016-03-03 · Investigation of Stabilization Methods for Multi-Dimensional Summation-by-parts Discretizations

28Hughes, T. J. R., Franca, L. P., and Mallet, M., “A New Finite Element Formulation for Computational Fluid Dynamics:I. Symmetric Forms of hte Compressible Euler and Navier-Stokes Equations and the Second Law of Theromdynamics,” ComputerMethods in Applied Mechanics and Engineering, 1985.

29Diosady, L. T. and Murman, S. M., chap. Higher-Order Methods for Compressible Turbulent Flows Using EntropyVariables, AIAA SciTech, American Institute of Aeronautics and Astronautics, Jan 2015.

30Douglas, J. and Dupont, T., “Interior Penalty Procedures for Elliptic and Parabolic Galerkin Methods,” ComputingMethods in Applied Sciences, edited by R. Glowinski and J. Lions, Vol. 58 of Lecture Notes in Physics, Springer BerlinHeidelberg, 1976, pp. 207–216.

31Hicken, J. E., Del Rey Fernandez, D. C., and Zingg, D. W., “Simultaneous Approximation Terms for MultidimensionalSummation-by-parts Operators,” Aviation 2016 (under review), Dallas, Texas, June 2016.

32Bezanson, J., Edelman, A., Karpinski, S., and Shah, V. B., “Julia: A Fresh Approach to Numerical Computing,”November 2014.

33Fike, J. A., Jongsma, S., Alonso, J. J., and van der Weide, E., “Optimization with Gradient and Hessian InformationCalculated Using Hyper-Dual Numbers,” AIAA, 2011.

34Davis, T. A., “An Unsymmetric-Pattern Multifrontal Method,” ACM Transactions on Mathematical Software, Vol. 30,No. 2, 2004.

35Gebremedhin, A. H., Manne, F., and Pothen, A., “What Color Is Your Jacobian? Graph Coloring for ComputingDerivatives,” SIAM Review , Vol. 47, No. 4, 2005.

36Seol, S., Smith, C. W., Ibanez, D. A., and Shephard, M. S., “A Parallel Unstructured Mesh Infrastructure,” HighPerformance Computing, Networking, Storage and Analysis (SCC), 2012 SC Companion:, Nov 2012, pp. 1124–1132.

37Mattsson, K., Svard, M., Carpenter, M., and Nordstrom, J., “High Order Accurate Computations for Unsteady Aerody-namics,” Computers and Fluids, Vol. 36, No. 3, 2007, pp. 636–649.

38Erlebacher, G., Hussaini, M. Y., and Shu, C.-W., “Interaction of a shock with a longitudinal vortex,” Journal of FluidMechanics, Vol. 337, Apr 1997, pp. 129–153.

13

Dow

nloa

ded

by J

ason

Hic

ken

on F

ebru

ary

8, 2

016

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

6-13

28


Recommended